principles of robot autonomy iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2....
TRANSCRIPT
![Page 1: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/1.jpg)
Principles of Robot Autonomy INon-parametric filtering
![Page 2: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/2.jpg)
Today’s lecture• Aim• Learn about non-parametric filters
• Readings• S. Thrun, W. Burgard, and D. Fox. Probabilistic robotics. MIT press, 2005.
Sections 3.1 – 3.4, 4.1, 4.3, 7.1
11/1/19 AA 274 | Lecture 16 2
![Page 3: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/3.jpg)
Instantiating the Bayes’ filter
• Tractable implementations of Bayes’ filter exploit structure and / or approximations; two main classes• Parametric filters: e.g., KF, EKF, UKF, etc. • Non parametric filters: e.g., histogram filter, particle filter, etc.
11/1/19 AA 274 | Lecture 16 3
![Page 4: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/4.jpg)
: convex regions forming a partition of state space (e.g., grid cell)
Histogram filter• Key idea: use discrete Bayes’ filter as an approximate inference tool
for continuous state spaces
• Step #1: histogram filters decompose a continuous space into finitely many bins
State space11/1/19 AA 274 | Lecture 16 4
![Page 5: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/5.jpg)
Example
11/1/19 AA 274 | Lecture 16 5
![Page 6: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/6.jpg)
Histogram filter
• Step #2: assign to each region 𝑥",$ a probability 𝑝",$; probabilities are then approximated according to a piecewise scheme
11/1/19 AA 274 | Lecture 16 6
![Page 7: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/7.jpg)
Histogram filter• Step #3: discretize motion and measurements models, i.e.,
1. Select mean state as representative state
2. Approximate measurement model
3. Approximate transition model
• Step #4: execute discrete Bayes’ filter with discretized probabilities
and
11/1/19 AA 274 | Lecture 16 7
![Page 8: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/8.jpg)
Histogram filter
11/1/19 AA 274 | Lecture 14 8
• Belief 𝑏𝑒𝑙(𝑥$)represented as pmf{𝑝",$}
• Then one car run the usual discrete Bayes’ filter
![Page 9: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/9.jpg)
Particle filter
• Key idea: represent posterior 𝑏𝑒𝑙(𝑥$) by a set of random samples
• Allows one to represent non-Gaussian distributions and handle nonlinear transformations in a direct way
11/1/19 AA 274 | Lecture 16 9
![Page 10: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/10.jpg)
Particle filter
• Samples of posterior distribution are called particles, denoted as
• A particle represents a hypothesis about what the true world state might be at time t• Ideally, particles should be distributed according to
• Matching exact only as 𝑀 → ∞, but 𝑀 ≈ 1000 usually good enough• A particle filter constructs the particle set from the particle set
recursively, with the goal of matching the distribution 𝑏𝑒𝑙(𝑥$)11/1/19 AA 274 | Lecture 16 10
![Page 11: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/11.jpg)
Particle filter: algorithm
• The temporary particle set represents the belief 𝑏𝑒𝑙 𝑥$• The particle set represents
the belief 𝑏𝑒𝑙(𝑥$)• Importance factors are used to
incorporate measurement 𝑧$ in the particle set• After resampling, particles are
(as 𝑀 → ∞) distributed as
𝑏𝑒𝑙(𝑥$45)
𝑏𝑒𝑙(𝑥$)
Prediction:𝑏𝑒𝑙(𝑥$)
Correction:𝑏𝑒𝑙(𝑥$)
Importance factor
11/1/19 AA 274 | Lecture 16 11
![Page 12: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/12.jpg)
More on resampling
Resampling
11/1/19 AA 274 | Lecture 16 12
![Page 13: Principles of Robot Autonomy Iasl.stanford.edu/aa274a/pdfs/lecture/lecture_16.pdf · 2021. 2. 10. · Sections 3.1 –3.4, 4.1, 4.3, 7.1 11/1/19 AA 274 | Lecture 16 2. ... nonlinear](https://reader035.vdocuments.net/reader035/viewer/2022071408/6100235639a66738e32a0c66/html5/thumbnails/13.jpg)
Next time
11/1/19 AA 274 | Lecture 16 13